(285g) Dynamic Modeling and Experimental Setup of a Pilot-scale Liquid-liquid Separator

Continuous separation of liquid-liquid dispersions in a gravity separator is essential in many technical processes such as extraction, water treatment, or recycling rare earths 1. The phase separation behavior of liquid-liquid dispersions is mainly determined by its composition and the drop size distribution2. Fluctuations in composition, e.g., salt concentration or pH, are typical for fermentation, metal recovery, and recycling processes. These fluctuations can change the phase separation behavior, particularly the coalescence behavior, flooding the separator with the disperse phase and ultimately resulting in process shut down3. Furthermore, residence times in technical liquid-liquid gravity separators amount up to an hour, leading to instabilities in conventional PID-control strategies4,5. Therefore, the operation of separators with varying compositions requires continuous monitoring and model-based control. The latter requires a model capturing the essential dynamics of the separator, which however can still not fully be described by mechanistic models. Therefore, we utilize a physics-informed neural network (PINN) approach that learns from available physical knowledge in the form of differential equations and experimental data to reliably describe the separator dynamics6. For this approach, we construct a pilot-scale experimental setup and derive a simplified dynamic model of the separator to design and train a physics-informed neural network of the separator.
The main objective of the PINN model is to describe the trajectory of the dense-packed zone height and outlet turbidity as a function of physical properties and operating conditions such as total volume flow, phase fraction, and the Sauter mean diameter of the dispersion. We extend a mechanistic dynamic liquid-liquid separator model from Backi et al. by the dense-packed zone and appropriate transport terms for the dense-packed zone to fulfill our model requirements7 . Based on that simulation model, we train a PINN model and do a sensitivity study regarding the network structure. Since the literature lacks experimental data on a continuous liquid-liquid separator, we constructed an experimental setup for a pilot-scale continuous liquid-liquid separator in our lab. In our setup, we investigate the influence of volume flows, Sauter mean diameter, salt concentration, and temperature on the dynamic behavior of a DN200 Separator. The Sauter mean diameter, height of the dense-packed zone and interfacial level in the separator are optically measured and evaluated by a Mask Recurrent Convolutional Neural Network (Mask R-CNN) that was retrained for our purposes8. Our experimental setup allows us to produce enough data to train and validate our PINN model, suggest model improvements, and test control strategies.
In this contribution, we present and discuss the results of modeling the dynamics of the liquid-liquid separator and the first experimental results from our pilot-scale setup. The dynamic model of the separator shows the trajectory of the height of the dense-packed zone and outlet turbidity as a function of operating conditions and physical properties. In the first step, we analyze under which operating conditions, such as total volume flow, phase fraction, and Sauter mean diameter, and physical properties, such as coalescence parameter9, the dense-packed zone floods the separator. In the second step, we illustrate the dynamics of the separator by applying step changes of operating conditions and coalescence parameters to access the remaining response time of the controller. In the last step, we investigate the required inputs and measurements for our PINN model.
In our experimental setup, we generate experimental data about the flooding of the separator and trajectories of the height of the dense-packed zone. The experimental data serve to validate the in-silico-determined flooding boundaries and trajectories. After successful validation and, if necessary, retraining the PINN model, we will provide a proof-of-concept of the control strategy to cancel out salt as a disturbance to the system.

Acknowledgement:
This work is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under the topic “Hybride Physik-Neuronales Netzwerk Softsensoren für den dynamischen Betrieb von Flüssig-flüssig-Trennprozessen” within the framework of the priority program “Maschinelles Lernen in der Verfahrenstechnik. Wissen trifft auf Daten: Interpretierbarkeit, Extrapolation, Verlässlichkeit, Vertrauen”.

Literature Cited
1. Pereira Neves H, Max Dias Ferreira G, Max Dias Ferreira G, Rodrigues de Lemos L, Dias Rodrigues G, Albis Leão V, Barbosa Mageste A. Liquid-liquid extraction of rare earth elements using systems that are more environmentally friendly: Advances, challenges and perspectives. Separation and Purification Technology. 2022;282:120064.
2. Steinhoff J, Charlafti E, Reinecke L, Kraume M, Bart H-J. Investigation and development of gravity separators with a standardized experimental setup. Can. J. Chem. Eng. 2020;98(1):384–393.
3. Kamp J, Villwock J, Kraume M. Drop coalescence in technical liquid/liquid applications: a review on experimental techniques and modeling approaches. Reviews in Chemical Engineering. 2017;33(1):1–47.
4. Henschke M. Dimensionierung liegender Flüssig-flüssig-Abscheider anhand diskontinuierlicher Absetzversuche. Zugl.: Aachen, Techn. Hochsch., Diss., 1994. Fortschritt-Berichte / VDI Reihe 3, Verfahrenstechnik. Nr. 379. Düsseldorf: VDI-Verl., 1995.
5. Fridman E. Introduction to Time-Delay Systems: Analysis and Control. Systems and Control. Cham: Springer International Publishing AG, 2014.
6. Cai S, Mao Z, Wang Z, Yin M, Karniadakis GE. Physics-informed neural networks (PINNs) for fluid mechanics: a review. Acta Mech. Sin. 2021;37(12):1727–1738.
7. Backi CJ, Emebu S, Skogestad S, Grimes BA. A simple modeling approach to control emulsion layers in gravity separators. In: 29th European Symposium on Computer Aided Process Engineering. Computer Aided Chemical Engineering. Elsevier, 2019:1159–1164.
8. Sibirtsev S, Zhai S, Jupke A. Mask R-CNN based droplet detection in liquid–liquid systems, Part 3: Model generalization for accurate processing performance independent of image quality. Chemical Engineering Research and Design. 2024;202:161–168.
9. Henschke M. Determination of a coalescence parameter from batch-settling experiments. Chemical Engineering Journal. 2002;85(2-3):369–378.